JT · CONVEX HULLS Convex sets were first studied by Brunn in 1887. Later, Minkowski added...
Transcript of JT · CONVEX HULLS Convex sets were first studied by Brunn in 1887. Later, Minkowski added...
A GENESIS FOR COMPACT CONVEX SETS
APPROVED:
Major Professor f
y&rvu# /£• ^Lnor Professor //
JT I rector of the Department of Mathematics
Dean of the Graduate School
A GENESIS FOR COMPACT CONVEX SETS
THESIS
Presented to the Graduate Council of the
North Texas State University in Partial
Fulfillment of the Requirements
For the Degree of
1 MASTER OF SCIENCE
By
Ronald D. Ferguson, B. A.
Denton, Texas
May, 1969
TABLE OF CONTENTS
?
INTRODUCTION iv
Chapter
I. CONVEX HULLS t. . 1
II. SUPPORT THEOREMS 16
III. EXTREMAL CHARACTERIZATION OF CONVEX SETS 26
IV. A PROPOSITIONAL SUBSET OF THE GENESIS
OF A SET 35
APPENDIX 42
BIBLIOGRAPHY. 43
xix
INTRODUCTION
This paper was written in response to the following
question: what conditions are sufficient to guarantee
that if a compact subset A of a topological linear space . t
L3 is not convex, then for every point x belonging to
the complement of A relative to the convex hull of A
there exists a line segment yz such that x belongs to
yz and y belongs to A and z belongs to A? Restated in •
the terminology of this paper the question may be given
as follows: what conditions may be imposed upon a
compact subset A of L3 to insure that A is braced?
The pursuit of this problem gave rise to the
concept of the genesis of a compact convex set. The
class of braced sets having convex hull A could be
considered as a propositional subset of the genesis
of A.
In general, the terminology and language of this
paper is that followed by Valentine (1) in his book on
convex sets. Terms that are newly introduced herein
are marked by an asterisk.
i v
CHAPTER I
CONVEX HULLS
Convex sets were first studied by Brunn in 1887.
Later, Minkowski added considerably to the body of
knowledge that had begun to accumulate around convex
sets. As a preliminary to the consideration of some
theorems related to convex sets the following definitions
are introduced:
Definition 1.1 Suppose S is a set of elements
with a collection of subsets called open sets which
satisfy the following axioms:
(a) The union of each collection of open sets
is an open set.
(b) The intersection of each finite collection
of open sets is an open set.
Then S together with its collection of open sets is
called a topological space. The collection of open sets
is called the topology of S. It can readily be deduced-
from (a) and (b) that <j> and S are open sets. Henceforth,
the better known properties of topological spaces will
be assumed.
Definition' 1.2 The sets C£S and D£S are
complementary iff CUD=S and CftD=<j>.
The complement of A relative to B is the set
denoted by B^A={x| xeB and x^A }. Obviously if C and
D are complementary, then C=S^D. C denotes the
complement of C, i.e. C=S^C=D.
Definition 1.3 A closed set in a topological
space S is the complement of an open set, i.e. A is
closed iff A=SM3, where B is open.
Definition 1.4 A linear space I over the field
of real numbers R is a space for which there exist two
binary operations, vector addition ( indicated by + )
from £*£ into I and scalar multiplication ( indicated
by juxtaposition ) from Rx£ into I which satisfy the
following axioms:
If x,y,z e I and a,3 e I then
(1) x+y e I (4) ax e I
(2) x+y=y+x (5) a(x+y)=ax+ay
(3) (x+y)+z=x+(y+z) (6) a($x) = (a3)x
(7) (a+$)x=ax+3x
(8) for every pair x,y e I, there exists a z e I
such that x+z=y
(9) if 1 is the identity of the real field R, then
lx=x for every x e I.
Number (8) implies the existence of an additive
identity for vector addition since x=y is a possible
pair. 0 will be used to denote the additive identity
and is synonomous with the origin of a vector space.
Definition 1.5 Let ll be a linear space with a
topology. Let x e I. Then an open set N in the topology
of £ such that x e N is called a neighborhood of x and
is denoted by N(x).
Definition 1.6 Let S be a set in I. If x e S
then x is an interior point of S iff there exists a
neighborhood N(x) of x such that N(x)^S. The set of
all interior points of S is called the interior of S
and is denoted by intS.
It is well known that intS=S iff S is an open set.
Definition' 1.7 Let S be a set in £ and x e t.
Then x is said to be a limit point of S iff for every
neighborhood N(x) of x, N (x) SS^x}^.
Definition 1.8 Let S be a set in £. Then S'
denotes the set of all limit points of S and is called
the derived set of S. SuS1 is called the closure of
S and is denoted by clS.
One may easily verify that clS is a closed set.
Definition 1.9 Let S be a set in I. The set
of points x e £ such that for every neighborhood N'(x)
of x, N(x)ftS (j) and N(x)ftS <f> is called the boundary of
S and is denoted by bdS.
4
A boundary point of S is said to be isolated if
it is not a limit point of S.
Definition 1.10 Let f be a function which maps
a subset S in a tbpological space % into a set S* of a
topological space £*. The function f is said to be
continuous at a point x e S iff for each neighborhood
V[f(x)] of f(x) in 5* there exists a neighborhood U(x)
of x in £ such that if y e SflU(x), then f(y) e V; i.e.,
f-(snu) £V.
Definition 1.11 If a linear space £ has a topology
then the topology is Hausdorff iff for each x e £,
y e H, x^y, there exist two disjoint open sets containing
x and y respectively.
Definition 1.12 A topological linear space L is
a linear space with a Hausdorff topology such that the
operations of vector addition, f(x,y)=x+y denoted by
(x+y)/(x,y), and scalar multiplication, f(a,x)=ax denoted
by (ax)/(a,x), where x e L , y e L , a e R are continuous
in all variables jointly.
Definition 1.13 If x e £, y e I, the line segment
xy joining x and-y is the set of all points of the form
otx+(l-a)y where a e R , 0<a<l.
Definition 1.14 A set is said to be star-shaped
relative to a point x e I iff for each point y e S
xy?S.
Definition 1.15 A topological linear space L
is locally star-shaped iff for each neighborhood U(0)
of the origin 0 of L, there exists a neighborhood V(0)
of 6 such that V(9)£U(0) and such that V(0) is star-shaped
relative to 0.
Definition 1.16 If A is a set in £ and if x e I,
then the set A+x={y+x| y e A } is called the translate
of A by x.
Definition 1.17 If A is a set in I and a e R
aA={ax| x e A } is called the scalar multiple of A by a.
Several well known theorems are tacitly assumed
in this paper in order to devote time to properties
and theorems whose consequences have a greater bearing
on the subject under consideration. So it is without
apology that the following theorem is offered without'
proof:
Theorem 1.1 The continuity of vector addition
and scalar multiplication implies that each translate
and non-zero scalar multiple of an ope« set is open. tJ
Theorem 1.2 A topological linear space L is
locally star-shaped and locally symmetric.
Proof: Let U be a neighborhood of the origin 0.
Now ax/(a,x) is continuous at (0,0), so that by the
definition of continuity there exists a neighborhood
V(0) of 0 and a number 6 >0 such that if |a|<6 and if
x e V(0), then ax e U(0). Immediately, aV?U whenever
|a|<6. Consider the class T={aV| |a|<6 }. Then
by definition of a topology and neighborhood uT is
a neighborhood of 8. Also, N= uT£U(0), and N is
star-shaped and'symmetric relative to 0.
Theorem 1.3 If Ln is an n-dimensional topological
linear space, then LD is linearly isomorphic with
Euclidean n-space En. This means there exists a n n one-to-one linear bicontmuous map of L onto E .
The immediate import of this theorem is that theorems
n n holding in L are immediately applicable to E and
vice-versa. For a proof of Theorem 1.3 see Valentine
(1, pp. 7-8) .
Definition 1.18 A set S££ is convex iff for each
pair of points x,y e S, xy^S.
Definition 1.19 A topological linear space L is
locally convex iff for each neighborhood U(6) of 6 there
exists a convex neighborhood V(0) of 0 such that V(8)£U(0),
An obvious alternative to Definition 1.18 is that
a set S i s convex iff it is star-shaped relative to
every point x e S. Similarlyf Definition 1.19 could be
altered to conform with Definition 1.15.
Definition 1.20 A convex set in L which has an
interior point is called a convex body.
Theorem 1.4 The intersection of any collection of
convex sets in a linear space £ is a convex set.
Proof: Let C be a collection of convex sets
in I. Consider fiC. Suppose QC has two points x and y.
Then x and y belong to every element of C. But this
implies that xy?;A for every A e C, so that xy£f2C. Then
ftC is convex. If QC does not have at least two points,
then the conclusion is obvious.
Definition 1.21 The convex hull of a set SsL is
the intersection of all convex sets in L containing S
and is denoted by convS.
S is convex iff convS=S.
Definition 1.22 A set S of a topological linear
space L is bounded iff for each neighborhood N(0) of 0
there exists a positive scalar multiple of N(0), aN(0),
such that S£aN(0).
Definition 1.23 If A is a subset of S where S
is a topological space, then A is compact iff every
open covering of A has a finite subcover. 'In other words,
if T is the topology of S, A is compact iff for every
class of open sets C£T such that A^uC, there exists a
finite subset B of C such that A£uB.
It is well known that A£Lnis compact iff A is
closed and bounded.
• Definition 1.24 A flat or linear variety of a
linear space SL is a translate of a linear subspace of
8
I. Two flats are parallel iff one is a translate of
the other; i.e., if F and F* are flats in I, then F and
F* are parallel iff there exists an x e I such that
F=F*+x. Further, the dimension of.a flat is the
cardinality of the basis of the corresponding parallel
linear subspace. A flat of dimension 1 is called a
line.
Any good book on linear algebra can provide
information concerning the basis and dimension of a
vector space.
Theorem 1.5 Each finite dimensional flat of a
topological linear space L is closed.
Proof: Suppose that V is a finite dimensional
flat of L. Assume, by way of contradiction, that
x e clVM#<f>. Then there exists a flat of minimal
dimension containing V and x. Let W be such a flat.
Describe a relative topology on W as follows: let
T={WftU | U e T*, where T* is the topology of L }.
Now if x is in the closure of V relatiye to W, Theorem 1.3 • i
implies that x e V, a contradiction. Thus, there exists
a neighborhood N(x) of x relative to W such that N(x)ftV=(j>.
But by definition of T there exists a neighborhood of x,
U (x) e T* such that U(x)ftW£N(x). Since V£W, U(x)ftV=<j>.
But then x £ clV. Thus clV'vV=({) and V is closed.
The next theorem relates a convex set to its line
segment subsets. For a proof see Valentine (1, p.16).
Theorem 1.6 Let S be a set . in the n-dimensional
linear space &n. Let be defined recursively as
follows:
Sx= u xy S.= u xy
y«i x ^ i " 1
y si-i in~l i n
Then = convS if in satisfies the inequality 2 <n+l<2 n. n Theorem 1.7 If A and B are two convex sets in £,
then conv(AuB)= u [aA+(l-a)B] = D. 0£a£l
Proof: If x e A, y e B, then ax+(l-a)y e conv(AuB)
when O^oKl. Thus D^conv(AuB) immediately. Now suppose
z e D, w e D. By definition of D there exist points
x,x* e A and y,y* e B such that z e xy, w e x*y*. But
xx*£A, yy*CB, so that conv(xx*uyy*)CD. Then zw is a
subset of conv(xx*uyy*) so that zw?;D. Thus D is convex
and (AoB)£D. Immediately, D=conv(A B).
Definition 1.25 Let SQl. The kernel K of S is
the set of all points z e i such that zx^S for every
x e S. In other words, the kernel K of S is the set of
all those points with respect to which S is star-shaped.
Notice that S is convex iff S=K where K is the kernel of
S. Also, note that K is itself a convex set follows
immediately from the definition of K, so that K is often
referred to as the convex kernel of S.
10
Definition 1.26 The interior of a set ScL
relative to the minimal flat containing it is denoted
by intvS. For a linear space SL, if x e I, y e I,
x^y, then intvxy denotes the set of all points of the
form ax+(l-a)y, where 0<a<l. If x e I, then intvx=x.
Definition 1.27 A point of x e S is a core point
of S iff for each point y el, with y^x, there exists a
point z e. intvxy such that xz£S. The set of all such
core points of S is called the cores.
Theorem 1.8 If S£L is star-shaped relative to a
point p, then the clS is star-shaped relative to p.
Proof: Let y e clS. Choose some point of yp,
say Ay+(1-A)p where 0<^£1. Now, Ax+(1-A) p/(x) is continous
in x, so that for each neighborhood U of Ay+(1-A)p in L
there exists a neighborhood V(y) of y such that
Ax+(1-A)p e U whenever x e V(y). Since y e clS, let
x e VfJS. Then Ax+(1-A)p e UftS. But this implies that
Ay+(1-A)p e clS. Then yp^clS.
Corollary 1.8 Let C be a convex set in L. Then
the closure of C is convex.
Proof: Let z e clC, w e clC. As in Theorem 1.8
Ax+(1-A)y/(x,y) is continuous, so that if some point of
zw is chosen, say Az+(1-A)w where 0<A<1, then, as in
Theorem 1.8, zw£clC, so that clC is convex.
11
1.8 implies immediately that
as a result cconvS£cl(convS).
Dsed convex set which contains S.
as cl (cojivS) ccconv. Therefore,
Theorem 1.9 If S is a set in a topological
linear space L, then the closed convex hull of S, denoted
by cconvS is the same as the closure of the convex hull
of S.
Proof: Corollary
cl(convS) is convex, and
Note that cconvS is a clc
Hence, convS^cconvS. Thi
cl(convS)=cconvS.
Theorem 1.10 Supbose S is a set in L and S is
star-shaped relative to ja, point p. If p e intK, where
K is the convex kernel 0]E S, and if y e clS, then
intvyp^intS.
Proof: Let.y e clS, and without loss of generality
let y=4>. Now all that is necessary is to show that
apcints for all 0<os<l since y=<f>. Since p e intK, there
exists a neighborhood V(<f>) of <f> such that p+V(<J>) £intK.
Then, for 3=-a/(l-a) , by Theorem 1.1, the set 3v() is
open. Since e clS, there exists a point u e $[V(c)>) ] QS.
Thus, there exists a point v e V(tj>) such that $v e S.
is an open set contained in S,
Also, x=(1-a)6v+a(p+V)=ap and
The set U=(1-a)Bv+a(p+V)
since v e S, p+V^intK.
ap e U?intS. Thus intvypsintS.
12
Corollary 1.10 Let C be a convex set in a linear
topological space L. If x e intC, y.e clC, then
intvxycintc.
Proof: Since intC=intK, Theorem 1.10 implies that
intvxy^intC immediately.
Theorem 1.11 Let S be a set in L and suppose
that S is star-shaped relative to a point p. Then, if
p e intS; the intS is star-shaped relative to p.
Proof: Let y e intS and without loss of generality
suppose that p=9, and choose x=ay, 0<a<l. Since
y e intS, there exists a neighborhood U(y) of y such that
U(y)£intS. Since p e K, aU(y)£S, and since aU(y) is open,
ay e aU(y)£intS, for 0<a<l. Thus yp£intS, and intS is
star-shaped relative to p.
Corollary 1.11 If S is a convex set in L, then
intC is convex.
Proof: The proof is immediate from Theorem 1.11
since C is star-shaped relative to every point in intC.
Theorem 1.12 If S^L, then intS^coreS.
• Proof: Suppose z e intS. Let U be a neighborhood
of z with U^intS. Let y e L^{z}. Since F(a)=z+a(y-z) is
continuous at a=0, there exists a 6 such that if 0<a<6, r S=3 *
then x=F(a) e U. Thus x e intvzy and xy£U£intS. Then
z e cores.
13
Theorem 1.13 If C is a convex set in I, then
coreC is convex.
Proof: Let x e coreC, y e coreC. Hence y e C
and C=K, where K is the convex kernel of C. Without
loss of generality assume that y=0. Then, let z=ax,
where (Ka<l. Consider any vector u where u^0. Since
x e coreC, there exists a constant <5>0 such that
x+gu £ C for 0<s3<6. C is star-shaped relative to y=0
so that a(x+3u)^C for 0<;3<6. But this implies that
ax e coreC and thus xy^coreC. Then coreC is convex.
Theorem 1.14 If C is a convex body in L, then
coreC=intC=int(clC).
Proof: Let y e coreC and x E intC. Since
y e coreC, there exists a point z such that y e intvxz
and such that zy£C. Hence, z-e clC. By Corollary 1.10
intvxz^intC, so that y e intC. Consequently, coreC^intC^
Theorem 1.12 implies that intC^coreC; hence, intC=coreC.
Now, let y £ int clC, x e intC. There exists a point
z £ clC such that y £ intvzx. Again by Corollary 1.10,
intvzx<;intC. Hence, y £ intC so that int clC^intC.
Since C^int clC, then int clC=intC.
Theorem 1.15 If S is an open set in L, then
convS is open.
Proof: Corollary 1.11 implies that int convS is
convex. Since S^convS, and Sft(bd convS) =<p, then
14
Spirit convS. But int cortvS is convex, so that the
previous sentence implies convS£int convS. Also,
int convS£convS, so that convS=int convS. Thus,
convS is open.
It should be pointed out that the convex hull
of a closed set need not be closed. As an example con-
sider the the set S in Euclidean two-space E2 with
rectangular coordinates (x,y) defined by
S={(x,y)| x2y2=l, 0<y<«> }
S is closed but convS={(x,y)| y<0 } is not closed.
CHAPTER BIBLIOGRAPHY
1. Valentine, Frederick A., Convex Sets, New York, McGraw-Hill Book Company, 1964.
15
CHAPTER II
SUPPORT THEOREMS
The concepts of 'hyperplanes and half-spaces
are very useful in the formulation and proof of various
theorems concerning convex sets. As soon shall become
apparent, hyperplanes are related to flats in a very
intimate manner. In order to define a hyperplane the
notion of a linear functional is introduced.
Definition 2.1 Recall that a real valued function
is often called a functional. A linear functional is
a function from & to R which is additive and homogeneous,
that is
f(x+y)=f(x)+f(y) x £ y e &
f(ax)=af(x) x £ a e r.
Definition 2.2 Suppose HC&, then H is said to be
a hyperplane iff there exists a non-identically zero
linear functional f and a real constant ot e r such that
H={ x e £| f(x)=a } denoted by H= [f ;<*] .
Theorem 2.1 A hyperplane is a flat.
Proof: Suppose that f is a linear functional
and that f(x)£o, x e I. Let H=[f:0t]. Since ffo, let
16
17
z £ £, with f(z)^0. Choose an arbitrary point q e It
and since f(z)j^0, define p= [-f (q)/f (z) ] (z)+y where
y e H. Since f is linear,
f (p)=f{ [ — f?(q) /f (z) 3 (z)+y}
f(p) =-f(q)+f(y) so that
f(p)+f(q)=f(p+q)—f(y)=a.
Thus p+q e H. But this statement together with the
definition of p implies that &=H+Rz. Consequently,
H is a flat.
Definition 2.3 If f is a linear functional and
if AC&, then f(A)>a means f(x)>a for every x e A.
Similar definitions may be made for inequalities that
are strict or reversed.
Definition 2.4 The hyperplane H=[f:a] bounds
the set A££ if either f(A)>a or f(A)<a holds.
Definition 2.5 The sets {x e H| f(x)>a },
{ x e £| f(x)<a }, { x e £ | f(x)!>a }, and { x £ 11 f(x)<ot }
are called half-spaces of I.
Axiom 2.1 A partial order on a set S is a subset
0 of SxS such that if (x,y) e 0, then (y,x) £ 0, and if
(x,y) e 0 and (y,z) e 0 then (x,z) e 0. Suppose that
S is a partially ordered set, then S contains at least
one maximal linearly ordered subset, where a linear order
is a partial order with the additional condition that
(x,y) i 0 implies that (y,x) e 0.
18
Examples of partially ordered relations are
less than or equal and set inclusion "5." Zorn's
maximal principle is useful in the proof of a number of
theorems concerning convex sets. It is equivalent to
the axiom of choice and also to the Hausdorff maximal
principle. For a proof of this equivalence and a
discussion of related topics see Kelley (1, pp. 31-36).
Zorn's Maximal Principle If T is a partially
ordered set and each linearly ordered subset of T has
an upper bound in T, then T contains at least one
maximal element.
Theorem 2.2 If A and B are disjoint convex sets
in a linear space £, then there exist complementary
convex sets C and D of £ such that A^C and B^D.
Proof: Consider the class P={(A^,B^)}, where
A^ and B^ are convex sets in £ such that A flBj_=(j), A^A^, (
B^B^. P is not empty since (A,B) e P. Now define a
partial order on P as follows:
(Ai,Bi)<(Aj,Bj) if A ^ A j and B^Bj.
The union of every linearly ordered subset of elements
in P belongs to P. By Zorn's Maximal Principle there
exists a maximal element (C,D) in P.
Now all that remains is to show that C=D. :
Suppose that p e £^(CuD) . By Theorem 1.7
conv(Cu{p})= u [aC+(1-a){p}] and also 0 < a < l
19
conv(Du{p})=0<u<;L[aD+(l-a) {p}] . Since (C,D) is maximal,
there exist points x e Dftconv(Cu{p}), y e Cftconv(Du{p}),
such that x fj: C and y £ D. But the above implies that
there exist points c e C, d e D such that x e intvcp,
y e intvdp. The points c, d, p determine a triangle'and,
like medians, dxQcy tf). Since dx^D, cy^C, then C£JD (f>,
a contradiction. Thus C=D. »
Theorem 2.3 If f and g are linear functionals in
I such that [f :a] = [g:3] , then there exists a constant X^O
such that f=Xg, a=X3-
Proof: If f or g is identically zero, the theorem
is trivial. Suppose f^O and g|0. If 3=0, choose
z e &Mg:3]. If 3^0/ choose z e [g: 3 3 - Thus g(z)?*0.
Let p e Si. Then there exists a t e [g:3] such that p=t+yz,
where y e R. Now, f(p-t)=yf(z), and g(p-t)=yg(z). Then
g(p-t)/g(z)=y so that f(p-t)=[f(z)/g(z)][g(p-t)]. Let
X=f(z)/g(z). Thus f (p) -f (t) = Xg(p)-Xg (t) , and hence
f(p)-a=Xg(p)-X3. Since the previous statement is true
for all p e l , then it is true for.p=z. Consequently,
a-X3=f(z)-Xg(z)=0. Hence a=X3« Immediately, f(p)=Xg(p).
Definition 2.6 A hyperplane H is said to support
a set S at a point x e S iff x e H and H bounds S.
Theorem 2.4 Suppose A and B are convex subsets of
a linear space £, and that coreB fj), A <J>, and AftcoreB=4>.
Then there exists a hyperplane H=[f:a] such that f(A)<a
and f(B)>a.
20
Note that in a finite-dimensional linear
topological space "core" corresponds, to "interior."
For a proof of Theorem 2.4, which is a fundamental
separation theorem, see Valentine (3, p. 24).
Theorem 2.5 Suppose that B is a convex body
in a topological linear space L, and suppose that F
is a flat in L such that FfiintB=<J>. Then there exists
a hyperplane H containing F which bounds B.
Proof: By Theorem 2.4 there exists a hyperplane
H*=[f:a] such that f(F)<ot and f(B)>a. Then there
exists a translate H=H*+x of H* such that F?H and
H bounds B.
Theorem 2.6 A hyperplane H=[f:a] in L bounds a
nonempty open set iff f is continuous and'f|0.
For a proof of Theorem 2.6 consult Valentine
( 3, pp.25-26).
Theorem 2.7 A hyperplane H=[f:a] is closed iff
f is continuous with f^O.
Proof: Suppose H is closed. Since H^L, there
exists a point x £ H and since H is closed, x $ clH so
that there exists a neighborhood N(x) of x such that
N(x)QH=<f>. Now by Theorem 2.6, f is continuous since H
bounds N(x).
Suppose that f is continuous with f£0. Then
{ x | f(x)<a } and { x | f(x)>a } are open convex sets and
H=L^[{x|f(x)<a}u{x|f(x)>a}] so that H=[f:a] is closed.
21
Theorem 2.8 Suppose that C is a closed convex
set in a locally convex space L and that x e I>C.
Then there exists a closed hyperplane H through x such
that HflC=4 .
Proof: IAC is a neighborhood of x. Then there
exists a convex neighborhood N(x) of x such that
N(x)CL^C and N(x)ftC=<}> since L is Hausdorff, locally
convex, and C is closed. The interior of N(x) <}>, so
that Theorem 2.4 implies the existence of a closed
hyperplane H* separating N(x) and C. Now consider the
translate H of H* such that x e H. HflC=4>.
Theorem 2.9 If C is a convex body in a topological
linear space L, then through each bour*dary point of C ,,
there passes a closed hyperplane of support.
Proof: If C=L the conclusion is trivial. Suppose
that C^L. Let x e bdC, and note that {x} is a convex
set. By Corollary 1.11 intC is convex. But Theorem 2.4
implies that there exists a hyperplane H* which separates
intC and {x}. Now consider a translate H of H* such that
x e H. Then H is a hyperplane of support at x.
Theorem 2.10 Suppose that S is a set in a
topological linear space L and that int convS <{>. A point
p e int convS iff each hyperplane H through p strictly
separates at least two points of S.
Proof: Suppose that p e int convS and that there ?
exists a hyperplane H=[f:a] such that p e H and H does
22
not separate at least two points of S. But this
statement means that all points of S belong to a closed
half-space determined by H, say {x|f(x)<.a}, since f is
continuous by Theorem 2.6. But now it is apparent that
p e bd convS, a contradiction. Thus each hyperplane
through p separates points of S.
Now suppose that each hyperplane through p
strictly separates at least two points of S. And further
suppose, by way of contradiction, that p ^ int convS.
By Theorem 2.5 and Theorem 2.7, since int convS <J>, there
exists a closed hyperplane through p bounding S, a
contradiction. Thus p e int convS.
Theorem 2.11 A line through the interior of a
compact convex body A intersects bdA twice.
Proof: Let p e intA tj). Since A is compact, A^L.
Consider a ray Q from p. A is bounded, but Q is
unbounded so that there exists some point y e (L'vA)QQ.
Now let Q={x e L | x=ay+(l-a)p, a>0 }. Consider line
segment py={ x e Q | O^a^l }. Order the elements of py
as follows: if w e py, z e py, where w=3y+(l-g)p and
z=Ay+ (1-A) p for some CK3, A<1, then z>w iff A>$. The
pyQA is bounded above by y. Then there exists a least
upper bound k for pyflA. Obviously, k e bdA and also
k e QfibdA. Similarly, for a ray Q*={x e L | x=ay+(l-a)p,
a<0} the same conclusion as before is obtained. So that
QuQ* is a line through p and QoQ* intersects bdA in at
23
least two points. In addition, to intersect in more
than two points contradicts the convexity of A.
Theorem 2.12 Let S be a closed set in a
topological linear space L, and suppose that intS <}>.
Then S is convex iff through each boundary point of S
there passes a hyperplane of support to S.
Proof: The necessity is provided immediately
by Theorem 2.9.
Now suppose that through each boundary point of
S there passes a hyperplane of support to S. If S=L,
then S is convex. Suppose S^L. Then let x e intS and
y e I/vStj). As in Theorem 2.11, there exists a z e bdS
such that z e xy. By hypothesis, there exists a
hyperplane of support H=[f:a] through z. y £ H, for if
y e H, then x e H since x, y, and z are colinear. But
x e H would contradict that H is a hyperplane of support.
Thus the closed half-space H+={x|f(x)>a} determined by
H and containing x contains S but not y where y is any
point in the complement of S. Let T be the class of
all such half-spaces that contain S but no point y
external to A. Thus, OT=S. By Theorem 1.4, S is convex
since each element of T is convex.
A finite dimensional normed linear space is sometimes
called a Minkowski space. Consult the Appendix for a more
formal definition of a Minkowski space. For a proof of
the following theorem see Valentine (3, p. 40).
24
Theorem 2.13 If S is a compact set in a Minkowski
n
space L , then convS is compact.
The closed convex hull of a compact set is also
compact in a Ban'ach space. A Banach space is a normed
linear space in which the Cauchy criterion is sufficient
for the convergence of sequences. For a proof of
Theorem 2.13 in Banach spaces see Taylor.(2).
•CHAPTER BIBLIOGRAPHY
1. Kelley, J. L. , General Topology, Princeton, N. J., D. Van Nostrand Company, Inc., 1955.
2. Taylor, A. E., Introduction to Functional Analysis, New York, John Wiley & Sons, Inc., 1958.
3. Valentine, Frederick A., Convex Sets, New York, McGraw-Hill Book Company, 1964.
25
CHAPTER III
EXTREMAL CHARACTERIZATION OF CONVEX SETS
Sufficient notions have now been introduced to
enable an examination of some sets which share a convex
hull, in particular, those sets which are minimal and
generate a given convex hull. The theorems in this
section culminate in extremal characterization of
convex sets including the Krein-Milman Theorem and
other related results.
Definition 3.1 A point x e bdC, where C is
convex arid Cc;L, is called an exposed point of C iff there
exists a hyperplane of support H to C through x such
that HflC={x}.
Rather than introduce a norm in order to define a
n
Minkowski space L , the following theorem is offered
without proof and only the comment that En is a Minkowski
space so that the theorem is intuitively correct. For
a proof consult Valentine (2, p. 52).
Theorem 3.1 Let C be a closed convex set in a
Minkowski space L2 (E2). Each compact connected portion
of the boundary of C which is not contained in a line
26
' 27
segment contains an exposed point of C. If C is a line
segment, it has two exposed points.
Definition 3.2 If S is a convex set in Z, then • *
a point x e S is an extreme point of S iff no non- ' 1
degenerate segment in S exists which contains x in its
relative interior.
Note that x is an extreme point of a convex set S
iff S^{x} is convex. Also, the above definitions imply
that the exposed points of S are extreme points of S.
Definition 3.3 A non-empty set M of a set K£L
is called an extremal subset of K iff x e K, y e K,
and Mf2intvxy <}> implies that {x}u{y}£M.
An extreme point of a set K is an extremal subset
of K consisting of just a single point.
Definition 3.4 If C is a compact convex set of
a topological linear space L, then extC denotes the set
of extreme points of C and expC denotes the set of
exposed points of C.
Definition 3.5 A real valued function is often
called a functional. Let p be a function defined on a
subset of Z. Then p is a convex functional defined on
a convex subset S of £, iff
p[ax+(1-a)y]=ap(x)+(1-a)p(y)
holds for all x, y e S and 0<a<l.
28
Definition 3.6* Let A bo a compact convex subset
of LU. Then the set denoted by gA={x | x£A and convx=A)
is called the genesis of A.
The elements of gA are called the generators of A.
The set y e gA such that y£x for every x e gA ( if such
a set exists ) is called the antecedent of A. If gA is •
considered to be partially ordered in the usual sense • <
of partially ordering sets, then the antecedent of A is
a minimal set analogous to Zorn's maximal element.
The following examples provide some clue as to
the nature of sets having antecedents:
Example 3.1 The antecedent of a closed line
segment is the set of its endpoints.
Example 3.2 The antecedent of a closed triangular
region is the set of its vertices. Further, the
antecedent of a solid polyhedron is the set of vertices
of the polyhedron.
Example 3.3 The antecedent of A={x c En: Ix-pl^r}
is the set {x e ,En: |x-p|=r}.
Example 3.4 Consider the open ball A where
A={x e En: |x-p|<r}. Although A is not closed and hence
not compact, it is informative to note that there are
infinitely many subsets of A which share A as a convex
hull, but no smallest such set. On the other hand, the
set B={(x,y) e E2| y>x2} is unbounded and thus not
29
compact but has a smallest set {(x,y) e E2| y=x 2H B
whose convex hull is B.
Notice that in examples 3.1 - 3.3 the antecedent
of each convex set was the set of exposed points of the
convex set. One might be led to conjecture that the
exposed points of a compact convex set in Ln is the
antecedent of the compact convex set. Example 3.5
quickly dispels this notion.
Example 3.5 Let A={x e E2: |x-s|=r} and p e E2^A.
Let B=conv(Au{p}). The antecedent of B is not the set
of exposed points of B. Let pw and pz be the tangents
from p to circle A where w, z e A. The antecedent of B
is the set {closed major arc zw}u{p}. But z and w are
not exposed points of B. They are, however, extreme
points of B.
From the definitions of expC and extC, it is
obvious that expCi;extC. But Example 3.5 clearly shows
that extC is not a subset of expC. However, both extC
and expC are subsets of bdC.
Theorem 3.2 If the antecedent of a compact
convex set A Ln exists, it is unique.
Proof: Suppose x is the antecedent of a compact
convex set A. Now suppose, by way of contradiction,
that there exists a y e gA such that y is an antecedent
for A and y^x* But by definition of antecedent x?y and
30
yCx. Then y=x, a contradiction. So that if the
antecedent of A exists, it is unique.
Theorem 3.3 If F is a family of extremal
subsets of a set K, then a non-empty intersection of
any subfamily of F is an extremal subset of K.
Proof: Let f^ e F, i e A. Choose x e K, y e K.
If ( Q f.) fJintvxy cj), then f. f2intvxy?*(j>, so that for ieA i i
each f•, {x}u{y}?f•, i e A. Hence, {x}u{y}£ Q f.. i i i^A 1
Definition 3.7 * Let S be a set. Let T be a
collection of subsets of S, then (S,T) is said to be
a restricted topological space iff for every non-empty
subcollection F<;T
(a) uF e T and
(b) if F is finite, ftF E T.
Theorem 3.4 The genesis of a compact convex set
having an antecedent is a restricted topological space.
Proof: Let F be a finite subset of gA where A
is a compact convex subset of Ln such that A has an
antecedent. Let x be the antecedent of A. Obviously,
xcy for every y e F. Then x£ftF£A. But, convx^conv(QF)
and conv(AF)£A. Now convx=A, so that conv(QF)=A. But
this implies that fiF e gA. A similar argument produces
that if F*CgA, then uF* e gA.
31
Theorem 3.5 If K is a non-empty compact set in
a locally convex topological linear space L, then K has
at least one extreme point.
Proof: Let F be the collection of all compact
extremal subsets of K. Then F is not empty since KCF.
Partially order the collection F by set inclusion.
Since K is compact, any non-empty linearly ordered
subset of F has a compact non-empty intersection, which
by Theorem 3.3( is an extremal subset of K. By Zorn's
Maximal Principle, the set F has a minimal element,
denoted by F*.
Now all that remains is to show that F* is an
extremal subset of K consisting of exactly one point
and thus an extreme point of K. Suppose, by way of
contradiction, that F* contains two distinct points
x and y. Now {x} is a closed convex set and y e IA>{X}.
By Theorem 2.8, there exists a closed hyperplane H such
that y e H and HJHx}=<f>. Then x £ H and F*ftHj*<j>. Since
F*QH is clearly an extremal subset .of K, and since
x £ F*ftH, the set F* is not minimal. Hence F* can
contain at most one point. This statement implies that
F*e K is an extreme point of K, since an extreme point
is an extremal set consisting of exactly one point.
Theorem 3.6 Let S be a compact convex set in a
locally convex topological linear space L. Then for each
closed hyperplane of support to S, HfiextSj <j>.
32
Proof: Let H be a closed hyperplane of support
to S. HQS <j>, where H=[f:a] and S is a subset of one of
the half-spaces determined by H, say H+. Since H is
closed, HQS is a non-empty, compact convex set and by
Theorem 3.5 HQS has an extreme point x. If y e S'v-H,
z e S, then f(y)<a, f(z)jca, where without loss of
generality H+={x e L| f(x)<a }. Since*f(x)=a, then
x £ intvyz. Also, if y e HQS, z e HQS, then since
x is an extreme point of SQH, x £ intvyz. Thus x is
an extreme point of S.
The Krein-Milman Theorem, which follows, is a
well known theorem that describes a convex set in terms
of its extreme points. For a different proof of
Theorem 3.7 see Krein and Milman (1).
Theorem 3.7 If S is a compact set in a locally
convex topological linear space L, then the closed
convex hull of S is the same as the closed convex hull
of the set of extreme points of S.
Proof: S Lnce extS£S, then cl conv extS is a
subset of the closed convex hull of S. Now it remains
to show that S£cl conv extS. Suppose, by way of
contradiction, that S is not a subset of cl conv extS,
and then choose x from S^cl conv extS cJ). Since L is
a locally convex space, Theorem 2.8 implies that there
exists a closed hyperplane H*=[f:B] through x, where
33
without loss of generality, f(y)<$ for y e cl conv extS.
But S is compact, so that a hyperplane H=[f:a], a>3,
exists such that HftS <|>, f(S)<a, and thus HQcl conv extS=<f>.
But Theorem 3.6 'implies that since H is closed, Ii
contains an extreme point of S. This is a contradiction.
Thus, it is false that S is not a subset of cl conv extS.
Then cl conv S= cl conv extS.
Theorem 3.7 may be restated as follows:
Theorem 3.8 Suppose that C is a compact convex
subset of a locally convex Hausdorff linear space L and
x is a subset of C. Then C is the closed convex hull
of x iff each extreme point of C lies in the closure of x.
A finite dimensional normed linear space Ln is
locally convex, so that Theorems 2.13, 3.7, and 3.8 now
imply that every compact convex subset of Ln has an
antecedent, namely the set of extreme points of the
compact convex set. Recall also that Ln is linearly
isomorphic to En. These statements then form the proof
of the following:
Theorem 3.9 If S is a compact convex set in a
finite dimensio-nal normed linear (Minkowski) space Ln
or En, then S has an antecedent.
CHAPTER BIBLIOGRAPHY
1. • Krein, M. G. and D. P. Milman, "On Extreme Points of Regular Convex Sets," Studia Mathematica, IX (1940), 133-138.
2. Valentine, Frederick A., Convex Sets, New York, McGraw-Hill Book Company, 1964.
34
CHAPTER IV
A PROPOSITIONAL SUBSET OF
THE GENESIS OF A SET
An advantage of defining the genesis of a compact
convex subset of Ln is to discuss some properties of
some of the propositional subsets of the genesis. As a
case in point, the following definitions l^ad to one
such subset of gA, where A£E3:
Definition 4.1* If X is a set property and A
is a compact convex subset of Ln, then PA(X) will be
used to denote the subset of gA consisting of those
elements having property X, i.e., PA(X)={y e gA|y has X}.
Definition 4.2* Let A be a subset of L, then the
set denoted by augA= u xy is the augmentation of A. xeA yeA
Definition 4.3 A maximal connected subset of a
set A in a topological space S is called a component of A.
A set A in L is simply connected iff each component
of the complement of A is unbounded.
Alternately, a set is simply connected iff any
closed curve within it can be deformed continuously to
a point of the set without leaving the set.
35
36
Definition 4.4* Let A£Ln, A not convex. Suppose
x e convA^A and that there exists points z and y
belonging to A such that x=az+(l-a)y for some 0<a<l,
then x is said to be braced relative to A. zy is an
A-relative brace of x. If every point of convA^A has
an A-relative brace, then A is said to be a braced set.
Definition 4.5 A crosscut of a set S^L is a
closed segment xy such that intvxy^intS and such that
x e bdS and y e bdS.
Theorem 4.1 If an open set K has no crosscuts,
then it i-s the complement of a convex set.
Proof: Let x, y £ K. If xy£2K <j>, then K has a
crosscut which is a subinterval of xy since K is open.
Hence xy K=(f> and K is convex.
Definition 4.6 If S is a set in a linear space I
and if V is a K-dimensional flat, then SflV is called a
K-dimensional section of S. If H is a hyperplane of I,
then SQH is called a plane section of S.
Definition 4.7 A continuum in a topological
linear space L is a compact connected set. A closed
connected set in a Minkowski space Ln is called a
generalized continuum . A generalized continuum is
automatically boundedly compact.
Definition 4.8* If A£L3 and A is not convex and
x e convA^A implies that there exists a plane H such
37
that x c H and AilH is connected and x e conv(AflH),
then A is said to be rigid.
Note that the set of braced sets whose convex
hull is convA is a propositional subset of g(convA).
If B is the property of being braced and C is a convex
set then The following lemma is presented
preparatory to determining a sufficient condition for
a set in L2 to be braced. Because of Theorem 1.3, a
proof in E2 is adequate.
Lemma 4.1 ; If A is a connected subset of E2
and p e A and q e augA^A and x=ap+(l-a)q for 0<a<l,
then x e augA.
Proof: If x e A, the result follows immediately.
Suppose x £ A. q e augA^A iff there exist 0<3<1 and
w, z e A such that q=f3w+(l-3)z. Let
[f:a]={y e E2| y=Ap+(l-A)q, A e R},
[g:S]={y e E2| y=Az+(l-A)x, A e R},
[h:y] = {y e E2 | y=Aw+(l-A)x, A e R} .
Note that if f(w)>a, then f(z)<a.
If g(p)>6, then g(q)<6, g(w)<6.
If h(p)>y, then h(z)<y, h(q)<y.
The above statements may be adopted without loss
of generality.
Clearly, p and w are elements of a connected set
A separated by Cgx 63 . Then D=AQ [g: otherwise A is
38
a subset of the union of the two open half-planes
determined by [g:6], contradicting that A is connected.
Case I: Suppose T^=Dft{x| f (x) Let t^ e T^.
Then there exists 0<A<1 such that x=Az+(1—A)t^. But
t^ e A, thus x e augA.
Case II: Suppose that T =4>. Then Dfi[f:a] (J) or
T2=Dfi{x | f (x) <a}j(j). But DQ[f:a] <{> implies that x e A,
a contradiction. Thus T2 <J>. Let t2 e T2" Then there
exists 0<A<1 such that t2=Az+(1-A)x. Thus h(t2)<Yr and
t2 and p are separated by [h:y].
Let £ T =Aft [h: y] ft{x | f (x) <a}fi{x | g (x) >6 . Then
t =Ax+(l-A)w where A>1. Now x=(1/A)t--t(1-A)/A]w, and 3
x=-[(l/A)-l]w+(l/A)t3. Let A=l/A. Then 0<A<1 so that
x=Xt^+(l-X)w. The last statement implies that
x e augA.
Theorem 4.2 If A is a compact connected subset
of E 2, then augA=convA.
Proof: Obviously augA<;convA.
Now suppose that x e convA. Since A is compact,
if x e bd convA, then x e extA^A or x belongs to the
relative interior of some maximal line segment zy where
zy^bd convA. Now line zy bounds A so that by Theorem 3.6
and Definition 3.2, z, y e extA^A. Thus if x e bd convA,
then x e augA.
39
Suppose that x e int convA. Let p be an extreme
point of A, thus p e bd convA. Consider the line
q=Ax+(1-A)p. qftint convA cj> so that by Theorem 2.11,
qflbd convA^{p} consists of exactly one point, say k.
Note that q=Ak+(l-A)p and for some 0<A<1, x=Ak+(l-A)p.
If k e A, then x e augA.
If k £ A, k e augA and by Lemma 4.1, x e augA.
Thus, in either case, convA^augA, so that augA
and convA are identical, i.e., augA=convA.
It is now obvious that if A is a compact connected
subset of L2, then augA is the convex hull of A.
Theorem 4.3 now follows immediately.
Theorem 4.3 If A is a compact connected subset
of L2 or E2, then A is braced.
Observe also that Theorem 4.2 together with
Theorem 1.7 implies the following:
Theorem 4.4 If A and B are disjoint subsets of
E2, and A is a compact connected set and B is a compact
connected set, then AUB is braced.
One should not be too hasty and conclude that
connectedness is sufficient for a compact set to be
braced in L3. The following example illustrates:
Example 4.1 Let A be a subset of E3 formed by
taking the union of three non-coplanar line segments
having a common point of origin. A is compact and
40
connected, but no point in the interior of the closed
solid tetrahedron which is A's convex hull has a
brace relative to A.
Other examples may be easily constructed to show
that connectedness is not necessary for.a set to be
braced in E2 or E3 and consequently L2 and L3.
For a proof of the following theorem consult
Valentine (1, pp. 91-93).
Theorem 4.5 If S is a compact set in L3 and if
each plane section of S is a simply connected continuum,
then S is convex. (The converse is obviously true.)
Theorem 4.6 If A is a subset of E3 or L3 and
A is rigid, then A is braced.
Proof: The theorem follows immediately from the
definition of rigid and Theorem 4.3.
The results of Chapter IV are now summarized in
terms of. Definition 4.1.
Theorem 4.7 Let B be the property of being braced.
Let C be the property of being connected. Let R be the
property of being rigid. Now suppose that X is a convex
subset of L2 and Y is a convex subset of L3, then
PX(CKPX(B) and Py(R)cPy(B).
Proof: The results are immediate from Theorem 4.2
and Theorem 4.5.
CHAPTER BIBLIOGRAPHY
1. Valentine, Frederick A., Convex Sets, New York, McGraw-Hill Book Company, 1964.
41
APPENDIX
Definition A.l A set S^i, star-shaped with
respect to the origin Q, is called linearly bounded iff
each line through Q intersects S in a line segment.
Definition A.2 Let S be an open set in L which
is star-shaped with respect to 0 and which is linearly
bounded. Then the Minkowski distance functional p is
a real valued function defined on L as follows:
p(x)=A>0 where \x*=x, x* e bdS, and ax*^intS, 0<a<l.
If x e S, then p(x)£l.
Definition A.3 A topological linear space L is
said to be normable if it is locally convex and if it
contains a non-empty bounded open set.
Definition A.4 A normable topological linear
space L with a norm (as for example, the Minkowski
distance functional) is called a normed linear space.
Definition A.5 A finite-dimensional normed '!
linear space is called a Minkowski space.
AO
BIBLIOGRAPHY
Books
Kelley, J. L., General Topology, Princeton, N. J., D. Van Nostrand Company, Inc., 1955.
Taylor, A. E., Introduction to Functional Analysis, New York, John Wiley & Sons, Inc., 1958.
Valentine, Frederick A., Convex Sets, New York, McGraw-Hill Book Company, 1964.
Articles
Krein, M. G. and D. P. Milman, "On Extreme Points of Regular Convex Sets," Studia Mathematica, IX (1940), 133-138.